Low-bit quantization reduces the inference cost of LLM agents but can compromise task success. SteerDistill is a quantization-aware distillation method where a full-precision teacher guides the student to revise failed responses and iteratively refines its guidance using feedback from student execution, providing locally verified behavioral corrections.
The chart will appear after repeat observations. The current metric comes from the source.
23 source votesReal observations only. History before source connection is not reconstructed.
The method can be useful for optimizing the performance of LLM agents with limited computational resources, maintaining accuracy without requiring full-precision models.
External memory enables LLM agents to improve from experience, but can also lead to memory-induced degradation, where an agent fails on tasks it could solve without memory. To address this, WikiTrace, an evidence-linked memory system, is introduced, allowing agents to verify each lesson against its original trajectory.
2 days agoDaily PapersReframe is a post-training framework that aligns the observation state, not just the output distribution, for agentic visual reasoning. The method includes GD-GRPO and OG-MOPD, which enhance the use of tool-returned observations and repair observations during rollouts.
3 days agoDaily PapersTactile sensing is essential for contact-rich dexterous manipulation, but commercial tactile sensors have limitations. This work presents LingBot-Touch, a tactile foundation model that unifies data from various sensor types, including piezoresistive, magnetic, and visuotactile sensors. The model is compatible with different policy models, enhancing manipulation capabilities.
3 days agoDaily PapersA2-MEAS, an embodied multi-agent framework using learned correction codes, enhances robot autonomy. The method achieves high performance on tasks like BEHAVIOR 2026 and RoboTwin 2.0, with significant improvements in task success and failure recovery.
6 days agoDaily PapersIntroduction of the Embodied First Exam, an evaluation framework where coding agents can choose their control approach. It includes 1,156 tasks across 35 suites on 47 robots, with three evaluation settings. The strongest agent achieved a 34% success rate in the Standard Setting.
3 days agoHacker NewsA collection of instructions for making foldable paper airplanes.
3 days ago